Research And Realization Of Face Feature Extraction And Recognition Algorithms

Face recognition is a frontier topic in the domains of artificial intelligence and pattern recognition, which involves multiple disciplines, including pattern recognition, image processing, neural network, physiology, psychology, etc. Compared with other ways of biometric recognition, face recognition has the advantages of non-contact, concealment and promptness. Meanwhile, face recognition involves many disciplines and is hard to realize. Thus, research on face recognition provides great value for the theory development. Furthermore, as an important biometric recognition method, face recognition has been applied to many occasions, such as criminal investigation, certificate verification, access control systems, automatic monitoring, etc.An automatic face recognition system usually comprises three parts, face detection module, feature extraction module and face recognition module. The research of this paper focuses on the two parts of feature extraction module and face recognition module. Firstly, this paper gives analysis on the theories of wavelet transform, frequency-domain transform, Support Vector Machine (SVM) classifier, and two-Dimensional Principal Component Analysis (2D PCA). Then, this paper carries on experiments in ORL database and Yale database, and finds the best wavelet function and sub-image selection mode for face recognition, and the optimum filtering function for raising recognition rates. The experiments also provides information about the best Discrete Fourier Transform (DCT) coefficient selection proportion, the optimum SVM penalty factor and kernel parameter, and the optimum component number of 2D PCA. Secondly, according to different steps in face recognition, this paper provides studies on three pre-processing methods, including space filtering, frequency filtering and an adjusting algorithm based on average value. This paper also gives comparison among three feature extraction methods, including wavelet transform, DCT and 2D PCA, and then presents the experimental results in ORL database and Yale database. Finally, the methods discussed in this paper are combined to obtain higher recognition rates.Combining the theory of wavelet transform, image transform, SVM classifiers and 2D PCA, this paper completes different experiments and reaches some valuable conclusions. With the methods presented in this paper, which combines the pre-processing of image transform and different feature extraction ways, the best recognition rate obtained in ORL database can be 98% with five training samples and five testing samples. In Yale database, the highest recognition rate obtained is 93.3% with five training samples and six testing samples.